TrOCR large Printed outputs only in CAPITAL letters..why?

Models
1

ld
ld863×349 3.51 KB

from transformers import TrOCRProcessor, VisionEncoderDecoderModel

from PIL import Image

processor = TrOCRProcessor.from_pretrained(r".\TrOCR")
model = VisionEncoderDecoderModel.from_pretrained(r".\TrOCR")
image = Image.open(r".\ld.png").convert(“RGB”)
pixel_values = processor(image, return_tensors=“pt”,attention_mask=True,Truncation=True).pixel_values
generated_ids = model.generate(pixel_values,pad_token_id=2,eos_token_id=2,max_new_tokens = 10000) #for open-end generation.

generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
generated_text

OUTPUT: ‘THIS IS I1 MEANS L & 1’

2

Hi,

This question also has been asked at the original repository: [TrOCR, Huggingface] Inference predictions always upper case? · Issue #875 · microsoft/unilm · GitHub.

3

Facing the same problem here as well. It should return the text as it is in the image not all upper case.